02.03 · UNIT 01 · PyTorch in One Hour: foundations · Lesson

Computation graphs

A computation graph draws a calculation as nodes and edges. Values flow from inputs to the loss.

PLAIN-LANGUAGE INTRODUCTION

What is this?

A computation graph draws a calculation as nodes and edges. Values flow from inputs to the loss.

One simple example

With x1=1.1, w1=2.2, and b=0, the net input is 2.42, sigmoid gives 0.9183, and the loss is tensor(0.0852).

Input tensors and parameter tensors for one example.

A loss value and a graph that records every step.

The same graph gives the path for backpropagation.

The graph does not compute the gradients by itself.

WORK THROUGH THE IDEA

A computation graph is a directed graph. Each node holds a value, and each edge passes a value.
For logistic regression, the forward pass computes z = x1*w1 + b, then a = sigmoid(z), then the loss.
The worked values are z = 2.42, a = 0.9183, and loss tensor(0.0852).
PyTorch builds the graph in the background when a tensor tracks gradients.
Common mistake: reading the graph as the gradient. Backpropagation walks the graph later.